2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length.
29 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug
unsigned int sysctl_sched_latency
= 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 const_debug
unsigned int sysctl_sched_nr_latency
= 20;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield
;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 25 msec, units: nanoseconds)
63 * This option delays the preemption effects of decoupled workloads
64 * and reduces their over-scheduling. Synchronous workloads will still
65 * have immediate wakeup/sleep latencies.
67 const_debug
unsigned int sysctl_sched_batch_wakeup_granularity
= 25000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 1 msec, units: nanoseconds)
73 * This option delays the preemption effects of decoupled workloads
74 * and reduces their over-scheduling. Synchronous workloads will still
75 * have immediate wakeup/sleep latencies.
77 const_debug
unsigned int sysctl_sched_wakeup_granularity
= 2000000UL;
79 /**************************************************************
80 * CFS operations on generic schedulable entities:
83 #ifdef CONFIG_FAIR_GROUP_SCHED
85 /* cpu runqueue to which this cfs_rq is attached */
86 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
91 /* An entity is a task if it doesn't "own" a runqueue */
92 #define entity_is_task(se) (!se->my_q)
94 #else /* CONFIG_FAIR_GROUP_SCHED */
96 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
98 return container_of(cfs_rq
, struct rq
, cfs
);
101 #define entity_is_task(se) 1
103 #endif /* CONFIG_FAIR_GROUP_SCHED */
105 static inline struct task_struct
*task_of(struct sched_entity
*se
)
107 return container_of(se
, struct task_struct
, se
);
111 /**************************************************************
112 * Scheduling class tree data structure manipulation methods:
116 max_vruntime(u64 min_vruntime
, u64 vruntime
)
118 s64 delta
= (s64
)(vruntime
- min_vruntime
);
120 min_vruntime
= vruntime
;
126 min_vruntime(u64 min_vruntime
, u64 vruntime
)
128 s64 delta
= (s64
)(vruntime
- min_vruntime
);
130 min_vruntime
= vruntime
;
136 entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
138 return se
->vruntime
- cfs_rq
->min_vruntime
;
142 * Enqueue an entity into the rb-tree:
145 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
147 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
148 struct rb_node
*parent
= NULL
;
149 struct sched_entity
*entry
;
150 s64 key
= entity_key(cfs_rq
, se
);
154 * Find the right place in the rbtree:
158 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
160 * We dont care about collisions. Nodes with
161 * the same key stay together.
163 if (key
< entity_key(cfs_rq
, entry
)) {
164 link
= &parent
->rb_left
;
166 link
= &parent
->rb_right
;
172 * Maintain a cache of leftmost tree entries (it is frequently
176 cfs_rq
->rb_leftmost
= &se
->run_node
;
178 rb_link_node(&se
->run_node
, parent
, link
);
179 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
183 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
185 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
186 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
188 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
191 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
193 return cfs_rq
->rb_leftmost
;
196 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
198 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
201 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
203 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
204 struct sched_entity
*se
= NULL
;
205 struct rb_node
*parent
;
209 se
= rb_entry(parent
, struct sched_entity
, run_node
);
210 link
= &parent
->rb_right
;
216 /**************************************************************
217 * Scheduling class statistics methods:
220 static u64
__sched_period(unsigned long nr_running
)
222 u64 period
= sysctl_sched_latency
;
223 unsigned long nr_latency
= sysctl_sched_nr_latency
;
225 if (unlikely(nr_running
> nr_latency
)) {
226 period
*= nr_running
;
227 do_div(period
, nr_latency
);
233 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
235 u64 period
= __sched_period(cfs_rq
->nr_running
);
237 period
*= se
->load
.weight
;
238 do_div(period
, cfs_rq
->load
.weight
);
243 static u64
__sched_vslice(unsigned long nr_running
)
245 unsigned long period
= sysctl_sched_latency
;
246 unsigned long nr_latency
= sysctl_sched_nr_latency
;
248 if (unlikely(nr_running
> nr_latency
))
249 nr_running
= nr_latency
;
251 period
/= nr_running
;
257 * Update the current task's runtime statistics. Skip current tasks that
258 * are not in our scheduling class.
261 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
262 unsigned long delta_exec
)
264 unsigned long delta_exec_weighted
;
267 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
269 curr
->sum_exec_runtime
+= delta_exec
;
270 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
271 delta_exec_weighted
= delta_exec
;
272 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
273 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
276 curr
->vruntime
+= delta_exec_weighted
;
279 * maintain cfs_rq->min_vruntime to be a monotonic increasing
280 * value tracking the leftmost vruntime in the tree.
282 if (first_fair(cfs_rq
)) {
283 vruntime
= min_vruntime(curr
->vruntime
,
284 __pick_next_entity(cfs_rq
)->vruntime
);
286 vruntime
= curr
->vruntime
;
288 cfs_rq
->min_vruntime
=
289 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
292 static void update_curr(struct cfs_rq
*cfs_rq
)
294 struct sched_entity
*curr
= cfs_rq
->curr
;
295 u64 now
= rq_of(cfs_rq
)->clock
;
296 unsigned long delta_exec
;
302 * Get the amount of time the current task was running
303 * since the last time we changed load (this cannot
304 * overflow on 32 bits):
306 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
308 __update_curr(cfs_rq
, curr
, delta_exec
);
309 curr
->exec_start
= now
;
313 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
315 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
318 static inline unsigned long
319 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
321 unsigned long weight
= se
->load
.weight
;
323 if (unlikely(weight
!= NICE_0_LOAD
))
324 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
330 * Task is being enqueued - update stats:
332 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
335 * Are we enqueueing a waiting task? (for current tasks
336 * a dequeue/enqueue event is a NOP)
338 if (se
!= cfs_rq
->curr
)
339 update_stats_wait_start(cfs_rq
, se
);
343 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
345 schedstat_set(se
->wait_max
, max(se
->wait_max
,
346 rq_of(cfs_rq
)->clock
- se
->wait_start
));
347 schedstat_set(se
->wait_start
, 0);
351 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
355 * Mark the end of the wait period if dequeueing a
358 if (se
!= cfs_rq
->curr
)
359 update_stats_wait_end(cfs_rq
, se
);
363 * We are picking a new current task - update its stats:
366 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
369 * We are starting a new run period:
371 se
->exec_start
= rq_of(cfs_rq
)->clock
;
375 * We are descheduling a task - update its stats:
378 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
383 /**************************************************
384 * Scheduling class queueing methods:
388 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
390 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
391 cfs_rq
->nr_running
++;
396 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
398 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
399 cfs_rq
->nr_running
--;
403 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
405 #ifdef CONFIG_SCHEDSTATS
406 if (se
->sleep_start
) {
407 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
412 if (unlikely(delta
> se
->sleep_max
))
413 se
->sleep_max
= delta
;
416 se
->sum_sleep_runtime
+= delta
;
418 if (se
->block_start
) {
419 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
424 if (unlikely(delta
> se
->block_max
))
425 se
->block_max
= delta
;
428 se
->sum_sleep_runtime
+= delta
;
431 * Blocking time is in units of nanosecs, so shift by 20 to
432 * get a milliseconds-range estimation of the amount of
433 * time that the task spent sleeping:
435 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
436 struct task_struct
*tsk
= task_of(se
);
438 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
445 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
447 #ifdef CONFIG_SCHED_DEBUG
448 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
453 if (d
> 3*sysctl_sched_latency
)
454 schedstat_inc(cfs_rq
, nr_spread_over
);
459 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
463 vruntime
= cfs_rq
->min_vruntime
;
465 if (sched_feat(USE_TREE_AVG
)) {
466 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
468 vruntime
+= last
->vruntime
;
471 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
472 vruntime
+= __sched_vslice(cfs_rq
->nr_running
)/2;
474 if (initial
&& sched_feat(START_DEBIT
))
475 vruntime
+= __sched_vslice(cfs_rq
->nr_running
+ 1);
478 if (sched_feat(NEW_FAIR_SLEEPERS
))
479 vruntime
-= sysctl_sched_latency
;
481 vruntime
= max_t(s64
, vruntime
, se
->vruntime
);
484 se
->vruntime
= vruntime
;
489 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
492 * Update the fair clock.
497 place_entity(cfs_rq
, se
, 0);
498 enqueue_sleeper(cfs_rq
, se
);
501 update_stats_enqueue(cfs_rq
, se
);
502 check_spread(cfs_rq
, se
);
503 if (se
!= cfs_rq
->curr
)
504 __enqueue_entity(cfs_rq
, se
);
505 account_entity_enqueue(cfs_rq
, se
);
509 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
511 update_stats_dequeue(cfs_rq
, se
);
513 #ifdef CONFIG_SCHEDSTATS
514 if (entity_is_task(se
)) {
515 struct task_struct
*tsk
= task_of(se
);
517 if (tsk
->state
& TASK_INTERRUPTIBLE
)
518 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
519 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
520 se
->block_start
= rq_of(cfs_rq
)->clock
;
525 if (se
!= cfs_rq
->curr
)
526 __dequeue_entity(cfs_rq
, se
);
527 account_entity_dequeue(cfs_rq
, se
);
531 * Preempt the current task with a newly woken task if needed:
534 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
536 unsigned long ideal_runtime
, delta_exec
;
538 ideal_runtime
= sched_slice(cfs_rq
, curr
);
539 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
540 if (delta_exec
> ideal_runtime
)
541 resched_task(rq_of(cfs_rq
)->curr
);
545 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
547 /* 'current' is not kept within the tree. */
550 * Any task has to be enqueued before it get to execute on
551 * a CPU. So account for the time it spent waiting on the
554 update_stats_wait_end(cfs_rq
, se
);
555 __dequeue_entity(cfs_rq
, se
);
558 update_stats_curr_start(cfs_rq
, se
);
560 #ifdef CONFIG_SCHEDSTATS
562 * Track our maximum slice length, if the CPU's load is at
563 * least twice that of our own weight (i.e. dont track it
564 * when there are only lesser-weight tasks around):
566 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
567 se
->slice_max
= max(se
->slice_max
,
568 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
571 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
574 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
576 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
578 set_next_entity(cfs_rq
, se
);
583 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
586 * If still on the runqueue then deactivate_task()
587 * was not called and update_curr() has to be done:
592 update_stats_curr_end(cfs_rq
, prev
);
594 check_spread(cfs_rq
, prev
);
596 update_stats_wait_start(cfs_rq
, prev
);
597 /* Put 'current' back into the tree. */
598 __enqueue_entity(cfs_rq
, prev
);
603 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
606 * Update run-time statistics of the 'current'.
610 if (cfs_rq
->nr_running
> 1)
611 check_preempt_tick(cfs_rq
, curr
);
614 /**************************************************
615 * CFS operations on tasks:
618 #ifdef CONFIG_FAIR_GROUP_SCHED
620 /* Walk up scheduling entities hierarchy */
621 #define for_each_sched_entity(se) \
622 for (; se; se = se->parent)
624 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
629 /* runqueue on which this entity is (to be) queued */
630 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
635 /* runqueue "owned" by this group */
636 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
641 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
642 * another cpu ('this_cpu')
644 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
646 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
649 /* Iterate thr' all leaf cfs_rq's on a runqueue */
650 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
651 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
653 /* Do the two (enqueued) entities belong to the same group ? */
655 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
657 if (se
->cfs_rq
== pse
->cfs_rq
)
663 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
668 #else /* CONFIG_FAIR_GROUP_SCHED */
670 #define for_each_sched_entity(se) \
671 for (; se; se = NULL)
673 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
675 return &task_rq(p
)->cfs
;
678 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
680 struct task_struct
*p
= task_of(se
);
681 struct rq
*rq
= task_rq(p
);
686 /* runqueue "owned" by this group */
687 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
692 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
694 return &cpu_rq(this_cpu
)->cfs
;
697 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
698 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
701 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
706 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
711 #endif /* CONFIG_FAIR_GROUP_SCHED */
714 * The enqueue_task method is called before nr_running is
715 * increased. Here we update the fair scheduling stats and
716 * then put the task into the rbtree:
718 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
720 struct cfs_rq
*cfs_rq
;
721 struct sched_entity
*se
= &p
->se
;
723 for_each_sched_entity(se
) {
726 cfs_rq
= cfs_rq_of(se
);
727 enqueue_entity(cfs_rq
, se
, wakeup
);
733 * The dequeue_task method is called before nr_running is
734 * decreased. We remove the task from the rbtree and
735 * update the fair scheduling stats:
737 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
739 struct cfs_rq
*cfs_rq
;
740 struct sched_entity
*se
= &p
->se
;
742 for_each_sched_entity(se
) {
743 cfs_rq
= cfs_rq_of(se
);
744 dequeue_entity(cfs_rq
, se
, sleep
);
745 /* Don't dequeue parent if it has other entities besides us */
746 if (cfs_rq
->load
.weight
)
753 * sched_yield() support is very simple - we dequeue and enqueue.
755 * If compat_yield is turned on then we requeue to the end of the tree.
757 static void yield_task_fair(struct rq
*rq
)
759 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
760 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
763 * Are we the only task in the tree?
765 if (unlikely(cfs_rq
->nr_running
== 1))
768 if (likely(!sysctl_sched_compat_yield
)) {
769 __update_rq_clock(rq
);
771 * Dequeue and enqueue the task to update its
772 * position within the tree:
779 * Find the rightmost entry in the rbtree:
781 rightmost
= __pick_last_entity(cfs_rq
);
783 * Already in the rightmost position?
785 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
789 * Minimally necessary key value to be last in the tree:
790 * Upon rescheduling, sched_class::put_prev_task() will place
791 * 'current' within the tree based on its new key value.
793 se
->vruntime
= rightmost
->vruntime
+ 1;
797 * Preempt the current task with a newly woken task if needed:
799 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
801 struct task_struct
*curr
= rq
->curr
;
802 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
803 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
806 if (unlikely(rt_prio(p
->prio
))) {
813 while (!is_same_group(se
, pse
)) {
814 se
= parent_entity(se
);
815 pse
= parent_entity(pse
);
818 delta
= se
->vruntime
- pse
->vruntime
;
820 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
824 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
826 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
827 struct sched_entity
*se
;
829 if (unlikely(!cfs_rq
->nr_running
))
833 se
= pick_next_entity(cfs_rq
);
834 cfs_rq
= group_cfs_rq(se
);
841 * Account for a descheduled task:
843 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
845 struct sched_entity
*se
= &prev
->se
;
846 struct cfs_rq
*cfs_rq
;
848 for_each_sched_entity(se
) {
849 cfs_rq
= cfs_rq_of(se
);
850 put_prev_entity(cfs_rq
, se
);
854 /**************************************************
855 * Fair scheduling class load-balancing methods:
859 * Load-balancing iterator. Note: while the runqueue stays locked
860 * during the whole iteration, the current task might be
861 * dequeued so the iterator has to be dequeue-safe. Here we
862 * achieve that by always pre-iterating before returning
865 static inline struct task_struct
*
866 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
868 struct task_struct
*p
;
873 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
874 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
879 static struct task_struct
*load_balance_start_fair(void *arg
)
881 struct cfs_rq
*cfs_rq
= arg
;
883 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
886 static struct task_struct
*load_balance_next_fair(void *arg
)
888 struct cfs_rq
*cfs_rq
= arg
;
890 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
893 #ifdef CONFIG_FAIR_GROUP_SCHED
894 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
896 struct sched_entity
*curr
;
897 struct task_struct
*p
;
899 if (!cfs_rq
->nr_running
)
904 curr
= __pick_next_entity(cfs_rq
);
913 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
914 unsigned long max_nr_move
, unsigned long max_load_move
,
915 struct sched_domain
*sd
, enum cpu_idle_type idle
,
916 int *all_pinned
, int *this_best_prio
)
918 struct cfs_rq
*busy_cfs_rq
;
919 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
920 long rem_load_move
= max_load_move
;
921 struct rq_iterator cfs_rq_iterator
;
923 cfs_rq_iterator
.start
= load_balance_start_fair
;
924 cfs_rq_iterator
.next
= load_balance_next_fair
;
926 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
927 #ifdef CONFIG_FAIR_GROUP_SCHED
928 struct cfs_rq
*this_cfs_rq
;
930 unsigned long maxload
;
932 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
934 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
935 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
939 /* Don't pull more than imbalance/2 */
941 maxload
= min(rem_load_move
, imbalance
);
943 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
945 # define maxload rem_load_move
947 /* pass busy_cfs_rq argument into
948 * load_balance_[start|next]_fair iterators
950 cfs_rq_iterator
.arg
= busy_cfs_rq
;
951 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
952 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
953 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
955 total_nr_moved
+= nr_moved
;
956 max_nr_move
-= nr_moved
;
957 rem_load_move
-= load_moved
;
959 if (max_nr_move
<= 0 || rem_load_move
<= 0)
963 return max_load_move
- rem_load_move
;
967 * scheduler tick hitting a task of our scheduling class:
969 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
971 struct cfs_rq
*cfs_rq
;
972 struct sched_entity
*se
= &curr
->se
;
974 for_each_sched_entity(se
) {
975 cfs_rq
= cfs_rq_of(se
);
976 entity_tick(cfs_rq
, se
);
980 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
983 * Share the fairness runtime between parent and child, thus the
984 * total amount of pressure for CPU stays equal - new tasks
985 * get a chance to run but frequent forkers are not allowed to
986 * monopolize the CPU. Note: the parent runqueue is locked,
987 * the child is not running yet.
989 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
991 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
992 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
994 sched_info_queued(p
);
997 place_entity(cfs_rq
, se
, 1);
999 if (sysctl_sched_child_runs_first
&&
1000 curr
->vruntime
< se
->vruntime
) {
1002 * Upon rescheduling, sched_class::put_prev_task() will place
1003 * 'current' within the tree based on its new key value.
1005 swap(curr
->vruntime
, se
->vruntime
);
1008 update_stats_enqueue(cfs_rq
, se
);
1009 check_spread(cfs_rq
, se
);
1010 check_spread(cfs_rq
, curr
);
1011 __enqueue_entity(cfs_rq
, se
);
1012 account_entity_enqueue(cfs_rq
, se
);
1013 resched_task(rq
->curr
);
1016 /* Account for a task changing its policy or group.
1018 * This routine is mostly called to set cfs_rq->curr field when a task
1019 * migrates between groups/classes.
1021 static void set_curr_task_fair(struct rq
*rq
)
1023 struct sched_entity
*se
= &rq
->curr
->se
;
1025 for_each_sched_entity(se
)
1026 set_next_entity(cfs_rq_of(se
), se
);
1030 * All the scheduling class methods:
1032 static const struct sched_class fair_sched_class
= {
1033 .next
= &idle_sched_class
,
1034 .enqueue_task
= enqueue_task_fair
,
1035 .dequeue_task
= dequeue_task_fair
,
1036 .yield_task
= yield_task_fair
,
1038 .check_preempt_curr
= check_preempt_wakeup
,
1040 .pick_next_task
= pick_next_task_fair
,
1041 .put_prev_task
= put_prev_task_fair
,
1043 .load_balance
= load_balance_fair
,
1045 .set_curr_task
= set_curr_task_fair
,
1046 .task_tick
= task_tick_fair
,
1047 .task_new
= task_new_fair
,
1050 #ifdef CONFIG_SCHED_DEBUG
1051 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1053 struct cfs_rq
*cfs_rq
;
1055 #ifdef CONFIG_FAIR_GROUP_SCHED
1056 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1058 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1059 print_cfs_rq(m
, cpu
, cfs_rq
);